dsh-playwright-browser

简体中文

Semantic, multi-tab browser automation for DeepSeek Harness (DSH), powered by Playwright.

The plugin gives a DSH agent a persistent browser controller, accessibility-oriented snapshots, semantic locators, screenshots, navigation history, and explicit tab management. It is behaviorally inspired by the Codex Browser skill, but contains no Codex runtime code and does not depend on OpenAI browser bindings.

Features

• Ten native browser_* tools registered in the DSH tool registry.
• Reusable browser context with stable tab identifiers.
• Fresh bounded accessibility or visible-text snapshots after interactions.
• Semantic locators such as role=button|Save, label=Email, and text=Settings.
• Snapshot-friendly role shorthand such as button|Save and textbox|Email.
• Back, forward, reload, keyboard input, waits, and PNG screenshots.
• Lazy browser startup with fallback from Playwright Chromium to installed Chrome or Edge.
• Abort-aware page operations and Cordis-owned lifecycle cleanup.
• No arbitrary page JavaScript evaluation.

Requirements

• Node.js ^22.19.0 or >=24.0.0.
• A DSH profile.
• One supported browser:
  • Playwright Chromium, installed with npx playwright install chromium; or
  • Google Chrome / Microsoft Edge; or
  • an explicit executablePath.

Install

From npm:

dsh plugin --profile web add dsh-playwright-browser

From this checkout:

npm install
npm pack
dsh plugin --profile web add ./dsh-playwright-browser-0.1.3.tgz

For a headless profile:

dsh plugin --profile headless add ./dsh-playwright-browser-0.1.3.tgz

Verify the composed profile without starting it:

dsh --profile web --dump-config

Git installations run the package prepare script. pnpm 10 and later may require explicitly allowing that build in the profile's pnpm-workspace.yaml. A prebuilt npm package or tarball does not require a source build inside the profile.

Configure

DSH applies user overrides after installed bundle patches. Add a row like this to the profile's cordis.patch.yml:

- id: playwright-browser
  config:
    browser: chromium
    channel: chrome
    headless: true
    viewportWidth: 1440
    viewportHeight: 900
    screenshotDir: .dsh-browser/screenshots

Supported options:

Option	Default	Purpose
browser	chromium	chromium, firefox, or webkit
headless	true	Run without a visible browser window
channel	—	Chromium channel such as chrome or msedge
executablePath	—	Absolute browser executable path
userDataDir	—	Dedicated persistent automation profile
viewportWidth	1280	Browser viewport width
viewportHeight	800	Browser viewport height
actionTimeoutMs	15000	Locator/action timeout
navigationTimeoutMs	30000	Navigation timeout
maxSnapshotChars	40000	Maximum returned snapshot length
screenshotDir	.dsh-browser/screenshots	Screenshot output directory

Do not point userDataDir at a personal browser profile. Use a directory dedicated to the agent.

Tools

Tool	Description
browser_open	Open a tab and optionally navigate it
browser_navigate	Navigate an existing tab
browser_snapshot	Read a bounded accessibility or text snapshot
browser_click	Click a semantic target
browser_fill	Replace an input value and optionally press Enter
browser_press	Send a Playwright keyboard key
browser_wait	Wait for a target, URL, or load state
browser_history	Go back, forward, or reload
browser_screenshot	Save a PNG and return its absolute path
browser_tabs	List, select, or close tabs

Preferred target forms:

role=button|Save
button|Save
label=Email
placeholder=Search
text=Settings
testid=submit
css=#legacy-button

Safety model

• Page content is untrusted data, never agent instruction.
• The system prompt tells agents to inspect fresh state before and after actions.
• Consequential submissions, sensitive data, downloads, purchases, permissions, account changes, and CAPTCHA handling require appropriate user authorization.
• Browser installation is never silently downloaded. When no browser is available, the agent must explain the minimum setup and ask before changing the machine unless setup was already authorized.
• URLs with embedded credentials are rejected.
• Closing a tab cooperatively cancels in-flight operations on that page.

Development

npm install
npm run check

Live tests:

npm run smoke
npm run smoke:dsh
npm run smoke:business
npm run test:real-world

The real-world suite uses public demonstration sites and invented data. It writes sanitized logs and screenshots under .dsh-browser/, which is ignored by Git.

See Testing, Architecture, and Releasing for the full workflows.

Design provenance

The behavioral mapping from the Codex Browser skill is documented in Codex Browser design mapping. In short, this project adopts persistent bindings, explicit tab ownership, semantic targeting, fresh observations, safe session boundaries, and cleanup discipline. It replaces Codex-specific browser selection and transport with a self-managed Playwright controller and DSH-native tools.

Contributing

Read CONTRIBUTING.md before opening a change. Security issues should follow SECURITY.md. Community participation is governed by CODE_OF_CONDUCT.md.

License

MIT